> Unlike popular diffusion models, OmniGen features a very concise structure, comprising only two main components: a VAE and a transformer model, without any additional encoders.
> OmniGen supports arbitrarily interleaved text and image inputs as conditions to guide image generation, rather than text-only or image-only conditions.
> Additionally, we incorporate several classic computer vision tasks such as human pose estimation, edge detection, and image deblurring, thereby extending the model’s capability boundaries and enhancing its proficiency in complex image generation tasks.
This enables prompts for edits like:
"|image_1| Put a smile face on the note."
or
"The canny edge of the generated picture should look like: |image_1|"
> To train a robust unified model, we construct the first large-scale unified image generation dataset X2I, which unifies various tasks into one format.
The author updated their code a couple of days ago, and it runs smoothly on my end, producing results in about one minute. https://github.com/VectorSpaceLab/OmniGen
I think there might be an issue with this website; it doesn't seem to be their official site. It's recommended to use the official code and demo: https://vectorspacelab.github.io/OmniGen/
Took a few minutes to load, some assets download at less than 1kbps. The first 3 times I got a "Connection error" after 30s. The 4th time has now been running for 5m.
I think this type of capability will make a lot of image generation stuff obsolete eventually. In a year or two, 75%+ of what people do with ComfyUI workflows might be built into models.
Using a single model to unify all image generation tasks, including many computer vision tasks and visual language reasoning, could transform future image generation models. Although some capabilities, like text-to-image, aren't perfect, it's a significant advancement. The model's ability to integrate so many tasks with strong instruction-following skills is impressive. I'm excited about the broad impact OmniGen could have on future research.
This looks promising. I love how you can reference uploaded images with markup - this is exactly what the field needs more of. After spending the last two weeks generating thousands of album cover images using DALL-E and being generally disappointed with the results (especially with the variations feature of DALL-E 2), I'm excited to give this a try.
I am working on a API to generate avatars/profile pics based on a prompt. I tried looking for train my own model bt I think it's a titanic task and impossible to do it myself. Is my best solution use an external API and then crop the face for what was generated?
The simplest commercial product for finetuning your own model is probably Adobe firefly, although there’s no API access support yet. But there are cheap and only slightly more involved options like Replicate or Civit.ai. Replicate has solid API support.
You can use a few controlnet templates and then whatever model you like and consistently get the posture correct. The diffusion plugin for Krita is a great playground for exploring this.
I sure hope so - at the very least I will use it for tabletop illustrations instead of having to describe a party's scenario result - I can give them a character-accurate image showing their success (or epic lack thereof).
It’s not really consistent - or anymore consistent than, say, SDXL with IP adapter. Even in their example images the character they’ve input comes out wearing different clothes.
I would say we already had one of those. There's more hand crafted human made content available than anyone cares to read.
While this will enable a certain degree of more spam it will more importantly, on the positive side of things, democratize the creative process to those who want to tell a story in images but lack the skill and resources to churn it out traditionally.
We literally already had AI fake porn of Taylor Swift making the rounds a while ago. Prepare for women in public positions to face that kind of bullshit more frequently.
It's more an issue of indifference than trust. For instance, you can show Trump supporters any number of legitimate videos that depict Trump and his associates saying, doing, and promising all kinds of outrageous, offensive, and destructive things, and they won't care in the slightest. It's not that they don't trust the video, it's that they've been programmed not to care. The leader cannot fail.
That's the ultimate purpose of disinformation -- it's not to make you believe false things, it's to make you believe nothing.
So yes, AI fakery will contribute to that phenomenon on behalf of numerous bad actors, but it was always going to happen anyway. You don't need Hinton and Sutskever on your side if you have Aisles and Murdoch.
> So yes, AI fakery will contribute to that phenomenon on behalf of numerous bad actors, but it was always going to happen anyway.
That's like saying: "Yes, crime might increase, but we will always have crime anyway." What will happen anyway is irrelevant precisely because it happens anyway. What's relevant is the expected increase in media distrust once everything might be a fake.
I was able to clone the repo and run it locally, even on a Windows machine, with only minimal Python dependency grief. Takes about a minute to create or edit an image on a 4090.
It's pretty impressive so far. Image quality isn't mind-blowing, but the multi-modal aspects are almost disturbingly powerful.
From https://arxiv.org/html/2409.11340v1
> Unlike popular diffusion models, OmniGen features a very concise structure, comprising only two main components: a VAE and a transformer model, without any additional encoders.
> OmniGen supports arbitrarily interleaved text and image inputs as conditions to guide image generation, rather than text-only or image-only conditions.
> Additionally, we incorporate several classic computer vision tasks such as human pose estimation, edge detection, and image deblurring, thereby extending the model’s capability boundaries and enhancing its proficiency in complex image generation tasks.
This enables prompts for edits like: "|image_1| Put a smile face on the note." or "The canny edge of the generated picture should look like: |image_1|"
> To train a robust unified model, we construct the first large-scale unified image generation dataset X2I, which unifies various tasks into one format.
Not exactly. They mention starting from the VAE from Stable Diffusion XL and the Transformer from Phi3.
Looks like these LLMs can really be used for anything
Took a few minutes to load, some assets download at less than 1kbps. The first 3 times I got a "Connection error" after 30s. The 4th time has now been running for 5m.
[1]: https://huggingface.co/Shitao/OmniGen-v1
Check out:
https://replicate.com/blog/fine-tune-flux
It is expensive though- Flux dev images are like $0.035/image
While this will enable a certain degree of more spam it will more importantly, on the positive side of things, democratize the creative process to those who want to tell a story in images but lack the skill and resources to churn it out traditionally.
Transparent Image Layer Diffusion using Latent Transparency
https://arxiv.org/abs/2402.17113
https://github.com/lllyasviel/sd-forge-layerdiffuse
Or, if you need solid regions that overlap and mask out other regions, then generate objects over a chroma-keyable flat background.
That's the ultimate purpose of disinformation -- it's not to make you believe false things, it's to make you believe nothing.
So yes, AI fakery will contribute to that phenomenon on behalf of numerous bad actors, but it was always going to happen anyway. You don't need Hinton and Sutskever on your side if you have Aisles and Murdoch.
That's like saying: "Yes, crime might increase, but we will always have crime anyway." What will happen anyway is irrelevant precisely because it happens anyway. What's relevant is the expected increase in media distrust once everything might be a fake.
It's pretty impressive so far. Image quality isn't mind-blowing, but the multi-modal aspects are almost disturbingly powerful.
Not a lot of guardrails, either.